论文主要思想笔记（1）

人脸识别，3d人脸重建，人脸老化过程，VR，3d-print

1. main idea:

the key challenge of face recognition is to develop effective feature representations for reducing intra-personal variations while enlarging inter-personal differences.

algorithm:

1) face alignment(SDM to detect facial landmarks)

2) feature extraction(400 DeepID2 vectors are extracted by a total of 200 deep ConvNets, each of which is trained to extract two 160-dimensional DeepID2 vectors on one particular face patch. ——use forward-backward greedy algorithm to select 4000 vectors——use PCA to extract 180 vectors from 4000 vectors )

3) face verification(Joint Bayes model)

dataset: LFW

reference:

Sun Y, Wang X, Tang X. Deep Learning Face Representation by Joint Identification-Verification[J]. Advances in Neural Information Processing Systems, 2014, 27:1988-1996.

1. main idea:

reconstruct 3D human face from large unstructured photo collections.

algorithm:

1) face landmark detection(find fiducial(landmarks) in each image, initialise shape to template)

2) pose estimation(solve for pose, warp each image to frontal using current shape estimate)

3) mask the images(solve for initial shape and lighting using photometric stereo; apply local view selection to re-estimate normals; resolve GBR(xyz of normal vector) and integrate a shape)

util convergence, goto step2.

dataset: LFW

reference:

Kemelmachershlizerman I, Seitz S M. Face reconstruction in the wild[C]// International Conference on Computer Vision. IEEE Computer Society, 2011:1746-1753.

1. main idea

facial aging simulator——physical model approaches

algorithm:

1) Geometry ageing(only change the width)

2) Texture ageing(by establish texture dictionary for each generation)

dataset: 3D faces(239 male / 218 female), private

reference:

Tazoe Y, Gohara H, Maejima A, et al. Facial aging simulator considering geometry and patch-tiled texture[J]. 2012:1-1.

1. main idea:

face aging model —— physical model approaches

algorithm: 

1) short term aging model(extend AAM model by adding muscles, then learn features in every part)

2) long term aging model(Markov Process)

dataset: FG-NET and MORPH

reference:

Suo J, Chen X, Shan S, et al. A concatenational graph evolution aging model.[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2012, 34(34):2083-96.

1. main idea

transform facial textures along perceived dimensions(such as age, gender, race, or health)——prototype approaches

algorithm:

1) construct a shape and texture prototype

2) subject + ( destination - source ) = new

age: transform18 to 45 or over 45 years old 

weakness: discard the personalised information and all the people share the same ageing pattern

reference:

Tiddeman B, Burt M, Perrett D. Prototyping and Transforming Facial Textures for Perception Research[J]. IEEE Computer Graphics & Applications, 2001, 21(5):42-50.

1. main idea:

3D colour printing. previous method: translucent. this method: full colour.

algorithm: an error diffusion half toning approach. 

1) voxelization: map color to surface voxels, generate slices

2) color management: gamut mapping, convert colours to tonal values

3) layer construction: push tonal values inside and extract layers

4) half-toning: consistent traversal; error diff. + material assign.

reference:

Brunton A, Arikan C A, Urban P. Pushing the Limits of 3D Color Printing: Error Diffusion with Translucent Materials[J]. Revista Mexicana De Astronomía Y Astrofísica, 2015, 12(12):180-180.

1. main idea:

near and far focus for one scene.

they use an optimisation algorithm(PDHG algorithm) to compute the images to be displayed on the presentation planes so that the retinal images when accommodating to different distances match the corresponding retinal images of the input scene as closely as possible

reference:

Narain R, Albert R A, Bulbul A, et al. Optimal presentation of imagery with focus cues on multi-plane displays[J]. Acm Transactions on Graphics, 2015, 34(4):1-12.

1. main idea:

they propose a system to match a given pair of virtual and physical worlds for immersive VR navigation.

design maps that are globally surjective to allow proper folding of large virtual scenes into smaller real scenes but locally injective to avoid locomotion ambiguity and intersecting virtual objects

aim: properly warp the virtual world appearance into real world geometry with sufficient quality and performance.

reference:

Sun Q, Wei L Y, Kaufman A. Mapping virtual and physical reality[J]. Acm Transactions on Graphics, 2016, 35(4).

1. main idea:

visual question answer(given an image, a question, after applying the algorithm, we can get the answer for the question according to the image)

algorithm:

1) question —— LSTM / PARSER——neural module networks——answer

2) image——CNN——neural module networks——answer

reference:

Jacob Andreas, Marcus Rohrbach, Trevor Darrell, et al. Neural Module Networks[J]. Computer Science, 2016, 27:55-56.

1. main idea:

add virtual objects to the real world, then they find that:

1) modern deep learning algorithms pre-trained on real data behave similarly in real and virtual worlds 

2) pre-training on virtual data improves performance 

reference:

Gaidon A, Wang Q, Cabon Y, et al. Virtual Worlds as Proxy for Multi-Object Tracking Analysis[J]. 2016.